🤖 Model Cost Advisor
Pick the most cost-effective LLM for any task — before you start spending.
Why pay Claude Opus prices for a task DeepSeek can handle? This skill analyzes your task, maps it to a capability tier, and finds the cheapest model that gets the job done well.
Quick Start
# 1. Fetch live pricing (one-time, auto-cached for 48h)
python scripts/fetch_pricing.py
# 2. Get a recommendation
echo "Write a REST API with FastAPI, handle auth and rate limiting" | python scripts/advise.py
# 3. Or pass task directly
python scripts/advise.py --task "Refactor a 2000-line Python class into smaller modules"
# 4. Compare all models side-by-side
python scripts/advise.py --compare
# 5. JSON output for scripting
python scripts/advise.py --task "Debug a race condition" --json
What It Does
- Analyzes your task description for complexity signals (reasoning depth, code needs, context length, agentic loops, domain expertise)
- Maps to one of 4 capability tiers: Budget → Standard → Advanced → Premium
- Estimates token usage based on task complexity
- Scores 30+ models using live pricing from litellm's community DB
- Recommends the top 3 models with projected cost, rationale, and pitfalls
The Four Tiers
| Tier | When to Use | Example Tasks | Typical Cost |
|---|
| 💰 Budget | Simple Q&A, classification, formatting, basic scripts | "Summarize this text", "Format JSON" | <$0.01 |
| 📦 Standard | Multi-step reasoning, medium code, structured output | "Write a web scraper", "Explain a concept" | $0.01–$0.10 |
| 🚀 Advanced | Complex code, architecture design, agentic loops | "Build a full-stack app", "Debug concurrency" | $0.10–$1.00 |
| 👑 Premium | Frontier reasoning, research, >128K context | "Research paper analysis", "Safety-critical code" | $1.00+ |
Models Tracked
30+ models across 6 providers, updated from litellm's community DB:
| Provider | Models |
|---|
| Anthropic | Claude Opus 4 / 4.1 / 4.5 / 4.6 / 4.7, Sonnet 4 / 4.5 / 4.6, Haiku 3.5 |
| OpenAI | GPT-4o, GPT-4o-mini, GPT-4.1 / 4.1-mini / 4.1-nano, o3 / o3-mini / o4-mini |
| Google | Gemini 2.0 Flash, 2.5 Flash / Pro |
| DeepSeek | V3 / V3.1 / V3.2, R1 (with reasoning token warning) |
| Alibaba | Qwen Turbo / Plus / Max / Coder-Plus / 3-235B |
| Mistral | Ministral 3B / 8B / 14B |
Example Output
╔══════════════════════════════════════════════════╗
║ 🤖 Model Cost Advisor ║
╚══════════════════════════════════════════════════╝
🎯 Task Analysis
Complexity Tier: 3 (Advanced)
Est. Input: ~24K tokens
Est. Output: ~10K tokens
Signals: multi_step_logic, complex_code, multi_turn_tools
💰 Top Recommendations
Rank Model Cost Input $/M Output $/M
───── ────────────────────── ──────── ──────── ─────────
🥇 deepseek-v3 $0.0175 0.28 0.42
🥈 deepseek-v3.1 $0.0216 0.27 1.10
🥉 gemini-2.5-flash $0.0322 0.30 2.50
📋 Why deepseek-v3?
Tier 3 task → best value in tier 1
Estimated total cost: $0.0175
How the Agent Uses This Skill
When loaded by Hermes, the agent follows these steps:
Step 1: Analyze Task Requirements
Classify the task along these dimensions to determine the minimum capability tier needed:
| Dimension | Weight | What to Assess |
|---|
| Reasoning Depth | High | Simple lookup → multi-step logic → deep chain-of-thought |
| Code Generation | Medium | None → simple scripts → multi-file complex → architecture design |
| Context Length | Medium | <4K → 4K-32K → 32K-128K → >128K tokens |
| Tool Use / Agentic | High | Single shot → multi-turn tools → autonomous agent loop |
| Domain Expertise | Low | General → specialized (math, legal, medical, Chinese content) |
| Output Quality | Medium | Draft OK → production → customer-facing critical |
| Latency | Low | Batch OK → real-time interactive |
Step 2: Estimate Token Usage
| Task Complexity | Input Tokens | Output Tokens |
|---|
| Trivial (single Q&A) | 500 – 2K | 200 – 1K |
| Simple (few exchanges) | 2K – 8K | 1K – 4K |
| Medium (multi-turn agent, 5-10 tools) | 8K – 40K | 4K – 16K |
| Complex (deep agent, 10-30 tools) | 40K – 150K | 16K – 50K |
| Heavy (autonomous loop, 30+ tools) | 150K – 500K+ | 50K – 200K+ |
Step 3: Run Scripts
# Ensure pricing is fresh
python scripts/fetch_pricing.py
# Get recommendation
python scripts/advise.py --task "<user's task description>"
Step 4: Present Recommendation
Format the output with:
- Task complexity analysis
- Top 3 model picks with cost
- Comparison vs user's current model (if known)
- Any pitfalls (R1 reasoning tokens, context window limits, etc.)
Pitfalls to Warn Users About
Script internals (for maintenance):
- Tier keys in pricing JSON are strings, not ints —
pricing_cache dict uses "1" not 1. The advise script casts them internally, but direct lookups must match.
- Keyword matching order matters — put longer-specific keywords (e.g.,
'production') before shorter ambiguous ones ('pr') to avoid substring false positives. Split on word boundaries.
User-facing pitfalls:
- R1/o3 reasoning tokens are hidden: Sticker price hides massive output consumption. Real cost is 3-5× higher for reasoning models.
- Context is not free: Models with 1M context (Gemini) charge for every token in the window, used or not.
- Tool calls compound cost: Every agentic round-trip adds system prompt + tool definitions + results. An agent task easily 5× the naive estimate.
- Cached prefixes save money: System prompts and cached prefixes bill at 10-25% — factor in for repetitive tasks.
- Chinese-language tasks: DeepSeek and Qwen outperform their price tier on Chinese content. Western models cost more for equivalent quality.
- Pricing changes frequently: Run
fetch_pricing.py before important decisions. Cache TTL is 48 hours.
Scripts
scripts/fetch_pricing.py — Fetches live pricing from litellm DB, normalizes to canonical model names, caches for 48h.scripts/advise.py — Task complexity analysis + model recommendation engine with colorized terminal output.
